【作者】 胡芝兰；

【导师】 林行刚；

【作者基本信息】 清华大学 ， 信息与通信工程， 2009， 博士

【摘要】 静态图像中的人体姿势估计是模式识别领域中的重要研究课题,也是许多计算机视觉系统的基础。目前,人体姿势估计技术逐步受到重视并取得了一定的进展。但是,复杂多变的背景、姿势和着装使得该问题仍为计算机视觉领域的一大难点。本文以人体上半身姿势为研究对象,从图像特征提取、部件检测、人体约束模型描述、以及姿势概率模型这一系列关键问题着手,提出了基于MF-OAR人体模型的上半身姿势估计算法,主要成果包括:第一,提出了基于Graph cuts的肤色分割算法。该方法有效结合了肤色、背景颜色以及像素的空间位置,并根据人脸实现前、背景颜色分布的动态建模,从而得到更为准确完整的肤色区域,为人体姿势估计提供了可靠的图像特征。第二,针对基于Graph cuts的图像分割技术的最小分割问题,提出了基于概率场的滤波思想:通过增强前、背景概率分布的可靠性,提高分割算法对光照及背景的鲁棒性。该滤波思想通过不同的实现方式在肤色区域分割和衣服分割中得到了有力的验证。第三,针对现有的基于单图像特征的部件检测方法产生大量不可靠候选的问题,提出了基于MF-OAR人体模型的关节点初始化方法:首先把高维姿势空间分解成三个独立子空间,并在子空间中完成对应关节点的初始化,计算复杂度低;同时,通过融合多种图像特征(边缘、前/背景颜色、肤色区域等)得到≤3个可靠初始候选,从而简化了姿势概率模型求解并提高姿势估计精度。第四,关节点初始化和全局姿势估计都为最大后验概率问题,其中的关键是似然函数的合理设计。本文设计的如下似然函数为最大后验概率地准确估计提供了保障。1)改进的方向性Chamfer匹配,提高了边缘似然函数对复杂背景、着装以及光照的鲁棒性;2)基于概率场的前/背景似然函数通过结合前、背景信息有效描述了每个姿势前景属性;3)肤色似然函数和区域似然函数通过合理结合肤色分割区域,实现了胳膊在姿势推理过程中的准确定位。本文提出的方法能够准确估计不同背景和着装下的多种姿势,在USC人体数据库中的估计性能明显优于相关算法,在本文收集的440幅图像中取得了62.1%的人体姿势估计成功率。更多还原

【Abstract】 Human pose estimation in static images is an important research topic in pattern recognition and the basis of many computer vision systems. Although there is lots of work in this field and some progresses have been made recently, recovering different poses in various backgrounds and clothing is still very challenging. To solve this problem, a novel upper pose estimation algorithm based on MF-OAR human model is proposed in this thesis, which includes image feature extraction, joints candidates detection, human pose constrain description, and pose probability model design. The main contributions are listed as follows.Firstly, a skin segmentation algorithm based on Graph cuts is proposed. It considers the skin color, background color and pixel location during the skin detection stage, with the skin color and background color being modeled dynamically based on the detected face. Compared with existing methods, it generates more accurate and complete skin regions for pose estimation.Secondly, a filter based on probability field is proposed to solve the smallest cut in the image segmentation methods based on Graph cuts. It improves the robustness of image segmentation methods against backgrounds and illuminations by enhancing the stability of foreground and background probability distributions. Implementations of such a filter in both skin region and clothing segmentations show that it improves the segmentation results greatly.Thirdly, most existing part detection methods find the human parts by using only one image feature, which produces a lot of poor part candidates. To solve this problem, a joints initialization method based on MF-OAR human model is proposed to get more accurate joints candidates. It firstly divides the pose space into three sub-spaces, and then each joint candidate is found within the corresponding sub-space to reduce the computation cost. Furthermore, it keeps up to three reliable candidates for each joint by integrating multiple image features (edge, foreground and background color, skin regions), which simplifies the pose inferring problem and gets more accurate pose estimation.Fourthly, joints initialization and pose estimation are both the problems of posterior probability maximization, where the key is selecting reasonable likelihood functions. In our method, the following likelihood functions are designed to improve the pose estimation capability: 1) the improved oriented Chamfer matching makes the edge likelihood more robust against background, clothing and illumination; 2) the foreground and background likelihood functions based on the probability field are proposed to determine whether each pose candidate is in the foreground or the background; 3) both the skin likelihood and region likelihood are designed by using the skin regions to help locating the arm during the pose inferring.Experiments on both the USC people database and 440 our collected images show that our method can recover upper body poses from images with a variety of individuals, poses, backgrounds and clothing more accurately than existing methods, with the success rate of 62.1% in 440 images.更多还原